⚡ electrical engineering

Assessing the Generalization of Graph Neural Networks for Fault Location Across Increasing Distributed Energy Resource Penetration Levels

This paper demonstrates that a Spatio-Temporal Graph Attention Network (STGATv2) significantly outperforms purely temporal, purely spatial, and traditional machine learning baselines in fault location across varying Distributed Energy Resource (DER) penetration levels, particularly by maintaining robust generalization and noise resilience when trained on high-penetration scenarios.

Original authors: Burak Karabulut, Olayiwola Arowolo, Carlo Manna, Chris Develder, Jochen L. Cremer

Published 2026-08-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Burak Karabulut, Olayiwola Arowolo, Carlo Manna, Chris Develder, Jochen L. Cremer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the electrical grid as a giant, invisible city of energy, where power flows like water through a complex network of pipes. For decades, this city was a one-way street: big power plants at the edge sent electricity straight to homes and factories. But today, the city is changing. Rooftop solar panels, backyard wind turbines, and battery packs are popping up everywhere, turning neighbors into mini-power plants. This is great for the planet, but it turns the grid into a chaotic two-way street. When a "short circuit" (a fault) happens—like a tree branch snapping a wire—the electricity doesn't just flow one way anymore; it bounces around from these new sources, making it incredibly hard for the grid's "doctors" to find exactly where the injury is. If they can't find the fault quickly, the lights stay out longer, and the whole city suffers. Scientists are trying to teach computers to be better detectives, using math that understands both the map of the pipes (spatial) and how the water rushes through them over time (temporal).

This paper is about teaching a new kind of computer detective, called a Spatio-Temporal Graph Neural Network (specifically a model named STGATv2), to find these electrical faults even when the grid is flooded with these new, unpredictable energy sources. The researchers simulated a standard electrical grid (the IEEE 123-bus feeder) and injected "faults" into it, then tested how well different computer models could spot the problem. They compared their new detective against older methods: one that only looked at the map (spatial), one that only looked at the timing of the rush (temporal), and some traditional math tricks. They tested these models under three different scenarios of "green energy" density: 10%, 25%, and 50% of the total load.

The results show that the new detective, STGATv2, is the clear winner. When the computer was trained on data that matched the test conditions, it got the answer right about 92% to 94% of the time. But the real magic happened when they tested its ability to adapt to changes it hadn't seen before. The researchers found that generalization is a one-way street. If you train the model on a grid with a lot of green energy (50% penetration), it stays smart even when the grid has less (10% or 25%). It's like a detective who learned to solve crimes in a chaotic, crowded city; they can still easily solve crimes in a quiet, empty town. However, if you train the model on a quiet grid (10% penetration) and then throw it into a chaotic one (50%), it gets confused. The older models (the ones that only looked at time or only looked at the map) crashed hard in this situation, with their accuracy dropping to the 69%–75% range. The STGATv2, however, held its ground, staying accurate at 81%–84% even in the chaos.

The paper also simulated "noise," which is like static on a radio or a foggy day for the sensors. Real-world measurements are never perfect. When they added this noise to the simulation, the model that only looked at time (called GRU) fell apart completely, dropping to as low as 33.5% accuracy in some cases. It was like a detective who couldn't hear a whisper over a loud fan. The STGATv2, which understands both the map and the timing, kept its cool, staying above 85% accuracy. This suggests that for the future of our power grids, the best way to find faults isn't just to watch the numbers change over time or just to look at the map; it's to have a detective that understands how the two work together, especially when the grid is full of unpredictable, renewable energy.

Technical Summary: Assessing the Generalization of Graph Neural Networks for Fault Location Across Increasing Distributed Energy Resource Penetration Levels

Problem Statement

Accurate fault location is fundamental to the reliability of power distribution networks. However, the increasing integration of Distributed Energy Resources (DERs) such as photovoltaics, wind, and battery storage complicates this task. DERs introduce intermittency, bidirectional power flows, and altered fault propagation patterns that reshape traditional fault signatures. While data-driven methods have emerged to address these challenges, existing literature often overlooks the underlying grid topology and the non-uniform electrical connectivity between buses. Furthermore, there is a lack of rigorous study regarding how Spatio-Temporal Graph Neural Networks (STGNNs) perform and generalize under varying and increasing levels of DER penetration, particularly when moving from low to high penetration scenarios.

Methodology

The authors propose a systematic evaluation of a Spatio-Temporal Graph Attention Network (STGATv2) designed to jointly model spatial and temporal dependencies in distribution grids.

Data Generation and Simulation:

Platform: Synthetic data was generated using OpenDSS and PyDSS on a reconfigured IEEE 123-bus feeder.
Configuration: The feeder was modified by opening the tie switch at (60, 160) and closing the one at (54, 94) to create more subtle voltage variations.
DER Penetration: Models were tested across three penetration levels (10%, 25%, and 50% of total load) using PV, wind, and BESS. Two placement strategies were employed: Localized (spatially biased) and Dispersed (uniform distribution).
Fault Scenarios: 25 fault locations were simulated with all 11 short-circuit types. Faults were characterized by moderate-to-high impedances (10–100 $\Omega$ ) and short durations (20 ms), resulting in subtle voltage drops (1–30 V).
Noise: Realistic measurement noise was introduced via zero-mean Gaussian perturbations at Signal-to-Noise Ratios (SNRs) of 50 dB and 45 dB.

Model Architecture:
The proposed pipeline (STGATv2) processes data in three stages:

Temporal Extraction: A Gated Recurrent Unit (GRU) processes per-phase RMS voltage time-series windows for each node to generate latent temporal representations.
Spatial Extraction: An improved Graph Attention Network (GATv2) layer aggregates these temporal embeddings. Unlike fixed-weight GNNs, GATv2 adaptively weights neighboring nodes based on their relative importance, capturing evolving fault signatures in active grids.
Classification: Node-level predictions are aggregated using soft voting to determine the final fault location or "no fault" status.

Baselines:
The study benchmarks STGATv2 against:

Purely Temporal: A standalone GRU model.
Purely Spatial: A GATv2 model where time windows are flattened into feature vectors.
Traditional ML: Random Forest (RF) and Support Vector Machines (SVM) with PCA.

Key Contributions

Systematic Benchmarking: The paper provides a quantitative comparison of spatio-temporal, purely spatial, and purely temporal models under multiple DER injection points and varying penetration levels.
Generalization Analysis: It rigorously evaluates model performance when training and testing across different DER penetration levels (e.g., training at 10% and testing at 50%), a scenario often unexplored in current literature.
Noise Robustness Assessment: The study isolates the impact of measurement noise on fault location accuracy, specifically examining how topological awareness aids in filtering noise under subtle fault conditions.

Results

In-Distribution Performance:

STGATv2 consistently outperformed all baselines, achieving macro F1 scores between 92% and 94% when training and testing at the same penetration level.
It significantly surpassed the purely spatial GATv2 (by 5–10 points) and the purely temporal GRU (by 8–19 points), indicating that joint spatio-temporal modeling is superior to modeling either dependency in isolation.

Generalization Across Penetration Levels:

Asymmetric Generalization: The study found that generalization is highly asymmetric. Models trained at high penetration (50%) retained near in-distribution performance when tested at lower levels (10% or 25%), with drops of at most 4 points.
Low-to-High Shift: Conversely, models trained at low penetration (10%) suffered significant degradation when tested at high penetration (50%).
- STGATv2 retained 81–84% F1 under these drastic shifts.
- GATv2 dropped to 69–74% F1.
- GRU dropped to 73–75% F1.
Traditional ML baselines (RF, SVM) degraded severely (dropping to ~47–60% F1) under unseen penetration levels, highlighting their inability to adapt to distribution shifts.

Impact of Measurement Noise:

Under 45 dB noise, STGATv2 maintained robust performance (>85% F1).
The purely temporal GRU model suffered the most, dropping to as low as 33.5% F1 in the dispersed configuration, demonstrating its inability to distinguish fault patterns from noise without topological context.
GATv2 performance also declined significantly (to ~74% F1), lacking the temporal memory to distinguish fault dynamics.

Significance and Claims

The paper claims that the integration of topological (spatial) awareness with temporal dynamics is critical for robust fault location in active distribution networks. Specifically:

Joint Modeling Superiority: The results confirm that jointly modeling spatial and temporal dependencies yields higher accuracy and robustness than approaches that treat them separately.
High-Penetration Training Strategy: The study suggests that training models on data with high DER penetration provides a diverse feature set that generalizes effectively to lower penetration levels, whereas low-penetration training fails to capture the complex bidirectional flows and fault signatures of high-penetration regimes.
Resilience to Noise: The authors conclude that topological awareness is fundamental for maintaining performance under realistic measurement noise, as it allows the model to filter out noise that masks subtle fault-induced voltage drops.

The authors position this work as a necessary step toward reliable grid operation in the face of increasing DER integration, noting that future work will focus on extending these frameworks to larger, diverse networks and evaluating extreme penetration regimes (up to 100%).